Smart Coverage Configurations: Recommender System for Car Insurance

Abstract

Car insurance policies in Zürich Insurance are heavily customizable as customers can choose between a number of coverages and adjust many options to fit their needs. The aim of the thesis is to develop a viable recommendation system solution which optimizes cover option proposals for motor insurance customers using adequacy metrics and feasible loss functions. The dataset provided by Zurich Insurance exhibited two major challenges, namely a high number of classes as well as a highly uneven distribution of number of occurrences per class. Before the model could be developed, the dataset needed to undergo pre-processing, structuring and dimension reduction. Two algorithms were chosen to develop the model: Multinomial Logistic Regression and XGBoost. The first algorithm showed to be challenging to model due to the high number of features. The second algorithm surpassed the model performance of the more traditional MLR model and produced an accuracy of 48.33%. The XGBoost model proved to be a suitable algorithm for the problem statement of Zurich Insurance. The model creates meaningful customer segmentation according to which coverages they purchased. Based on this segmentation it creates accurate recommendations for coverages and deductibles for new customers. The result is a better-informed customer who will not lose his/her time going through offers that do not meet his/her needs.